Spatial autocorrelation reduces model precision and

predictive power in deforestation analyses

KRISTJAN D. METS ,1,  DOLORS ARMENTERAS ,2 AND LILIANA M. DAVALOS

1,3

1
Department of Ecology and Evolution, Stony Brook University, Stony Brook, New York 11794 USA
2
Landscape Ecology and Ecosystem Modelling Laboratory, Department of Biology, Sciences Faculty,
Universidad Nacional de Colombia, Bogota 111321 Colombia
3
Consortium for Inter-Disciplinary Environmental Research, Stony Brook University, Stony Brook, New York 11794 USA

Citation: Mets, K. D., D. Armenteras, and L. M. Da
valos. 2017. Spatial autocorrelation reduces model precision and
predictive power in deforestation analyses. Ecosphere 8(5):e01824. 10.1002/ecs2.1824

Abstract. Generalized linear models are often used to identify covariates of landscape processes and to
model land-use change. Generalized linear models however, overlook the spatial component of land-use
data, and its effects on statistical inference. Spatial autocorrelation may artificially reduce variance in obser-
vations, and inflate the effect size of covariates. To uncover the consequences of overlooking this spatial
component, we tested both spatially explicit and non-spatial models of deforestation for Colombia. Param-
eter estimates, analyses of residual spatial autocorrelation, and Bayesian posterior predictive checks were
used to compare model performance. Significant residual correlation showed that non-spatial models
failed to adequately explain the spatial structure of the data. Posterior predictive checks revealed that spa-
tially explicit models had strong predictive power for the entire range of the response variable and only
failed to predict outliers, in contrast with non-spatial models, which lacked predictive power for all
response values. The predictive power of non-spatial models was especially low in regions away from
Colombia’s center, where about half the observations were clustered. While all analyses consistently identi-
fied a core of important covariates of deforestation rates, predictive modeling requires parameter estimates
informed by the spatial structure of the data. To inform increasingly important forest and carbon sequestra-
tion policy, land-use models must account for spatial autocorrelation.

Key words: Amazon; Andes; Colombia; land protection; posterior predictive check.

Received 29 July 2016; revised 31 January 2017; accepted 17 February 2017. Corresponding Editor: Tobias van Kooten.
Copyright: © 2017 Mets et al. This is an open access article under the terms of the Creative Commons Attribution
License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.
  E-mail: kristjan.mets@stonybrook.edu

INTRODUCTION Modeling spatial patterns under the assump-

Statistical testing and modeling lies at the core
of all analyses and projections of deforestation,
and land use, but statistical analyses generally
assume independence between observations
(Kruskal 1988). When data represent measures
taken across geographic space, the assumption of
independence is not fully met because of spatial
autocorrelation (Legendre 1993). Pairs of spatial
observations typically exhibit a correlation inverse
to the distance between them, and ecological spa-
tial data are no exception (Cliff and Ord 1970).
tion of independence affects statistical inference in
three ways. First, spatial autocorrelation emerges
as non-random geographic association of residual
errors in regression analyses (Cliff and Ord 1972).
Compared against spatially explicit models, then,
non-spatial models of spatial patterns have
deflated estimates of variance and residual auto-
correlation (Legendre and Fortin 1989). This leads
to loss of model precision and higher type I error
rates (Beale et al. 2010). Second, non-spatial mod-
els used for spatial data spuriously internalize
spatial autocorrelation into the goodness of fit of

❖ www.esajournals.org 1 May 2017 ❖ Volume 8(5) ❖ Article e01824


the model, undermining comparisons of model
performance (Telford and Birks 2005). Lastly,
when explanatory variables exhibit different spa-
tial patterns and degrees of autocorrelation, non-
spatially explicit models inflate the effect sizes of
the more autocorrelated variables (Lennon 2000).
Spatially explicit analyses are increasingly
used to address the problems arising from spatial
autocorrelation. Accounting for spatial autocor-
relation, however, remains relatively uncommon
in many disciplines, including the study of land-
use change (Brown et al. 2013). The increased
complexity these methods present to researchers,
as well as a common failure to recognize that
spatial autocorrelation in model residuals indi-
cates a violation of independence, prevents the
broad adaption of spatially explicit methods
(Ku€ hn and Dormann 2012). Within the study of
land-use change, the spatial data necessary to
reproduce land-use analyses are seldom pub-
lished along with the models. Most studies will
therefore never undergo critical re-analyses by
other researchers (Koenig 1999, Hunter et al.
2009). Hence, the consequences of assuming
independence of observations in models of land-
use change remain unknown.
Tropical deforestation is a prime example of an
intrinsically spatial process of change in land
use. Although analyses sometimes include a spa-
tial component, linear and generalized linear
regressions are two common approaches for
identifying relationships between deforestation
and a suite of explanatory socioeconomic, envi-
ronmental, and infrastructure variables (Rudel
and Roper 1997, Kaimowitz et al. 2004). With
few opportunities to control factors over time,
most analyses apply regressions against explana-
tory variables without regard to spatial structure.
Ignoring spatial autocorrelation makes these
analyses susceptible to bias from homogenous
and spatially concentrated data clusters (Over-
mars et al. 2003), even as the conclusions of these
analyses help shape national and international
efforts to curb deforestation by identifying and
discouraging enabling factors (Gullison et al.
2007). Improving the predictive power of defor-
estation analyses has gained urgency as these are
components of initiatives aimed at preserving
biodiversity or combating climate change such as
the Reducing Emissions from Deforestation and
Forest Degradation (REDD+) through the United
❖ www.esajournals.org
METS ET AL.
Nation’s Framework Convention on Climate
Change (UNFCCC), last renewed through the
2015 Paris Agreement (UNFCCC 2015).
Here, we present a series of non-spatially and
spatially explicit analyses to quantify the effects of
spatial structure when modeling deforestation
data. These models build on published data on
determinants of deforestation in Colombia
(Armenteras et al. 2011b, 2013a). We compare
parameter estimates and residual autocorrelation
from deforestation models that do and do not
explicitly capture spatial autocorrelation, and
evaluate the model performance through Baye-
sian posterior predictive checks. Bayesian poste-
rior predictive checks allowed us to produce
distributions of deforestation values for all munic-
ipalities based on estimated coefficients and com-
pare them to observed patterns of deforestation.
MATERIALS AND METHODS
Material and data
We reanalyzed change in forest cover for the
geographically heterogeneous country of Colom-
bia for 1985–2005 for the Andean region and
1990–2005 for the rest of the country (Armenteras
et al. 2011b, 2013a). Data for non-Andean forest
cover in 1990 and 2005 were obtained from the
report on “Scientific and institutional capacity
building to support REDD projects in Colombia”
(Montenegro et al. 2011). The original forest
cover dataset was measured via remote sensing
using over 240 Landsat multispectral satellite
images with <10% cloud obstruction from 1990
to 2005. For the Andean region, analyses by the
National University of Colombia provided the
land-cover data (Armenteras et al. 2011b). The
annual rate of deforestation of each municipality
(Rm, in %) was calculated as (Fearnside 1993):
Am1  Am2
Rm ¼  100
FA1  t
in which Am1 and Am2 are total forest areas
within the municipality at the beginning (1985 or
1990) and end (2005) of remote sensing records,
respectively. FA1 is the total forest area through-
out Colombia at the initial year, and t is the time
interval in years. Covariates of deforestation
included environmental and social variables
(Table 1) for 1119 municipalities divided into five
socio-environmental regions (Fig. 1).
2 May 2017 ❖ Volume 8(5) ❖ Article e01824
 METS ET AL.

Table 1. A summary of the potential covariates of forest cover change.

Variables Units Spatial resolution Description Source(s)

Urban population population/ha Municipality Change in urban population density DANE (1985, 1993, 2005)
density between 1985 and 2005
Rural population population/ha Municipality Change in rural population density DANE (1985, 1993, 2005)
density between 1985 and 2005
Unsatisfied basic percentage Municipality Percentage of population with DANE (2005)
needs (NBI) unsatisfied basic needs in 2005;
includes minimum household
connections, access to sanitary
services, access to primary education,
and minimum household economic
capacity as basic needs
Crops ha 30 m Change in crop area between 1985 and IDEAM (2007)
2005 measured via remote sensing
Pastures ha 30 m Change in pasture area calculated IDEAM (2007)
between 1985 and 2005
Illicit crops ha 10 m Area of coca (erythroxylum coca) crops UNODC (2006)
Cattle number Municipality Head of cattle per municipality in 2006 IGAC (2011)
Fire hotspots number Municipality Number of fire hotspots detected per NASA (2015)
municipality between 2000 and 2005
Mining kg Municipality Total gold and silver production in IGAC (2011)
2005
Protected area ha 1:100,000 Area under special management as IGAC (2005)
national protected area or indigenous
reserve
Road density km/ha 1:100,000 Density of roads in each municipality IGAC (2005)
Slope degrees 90 m Average maximum slope for each IGAC (2005)
municipality
Water scarcity index Municipality Index of water scarcity in a dry year IDEAM (2000)
Elevation ~m 90 m Altitude above sea level IGAC (2005)
Precipitation mm 1 km2 Total annual precipitation Worldclim (Hijmans et al. 2005)
Note: DANE, National Administrative Department of Statistics; UNDOC, United Nations Office on Drugs and Crime;
MADR, Agriculture and Rural Development Ministry; FIRMS, Fire Information for Resource Management System; SIGOT,
Geographic Information System for Planning and Territorial Development; IGAC, Agust
ın Codazzi Geographical Institute;
IDEAM, Institute of Meteorology and Environmental Studies, Worldclim global climate data.

Modeling approaches for spatial autocorrelation; (2) Bayesian random

Previously published analyses using the same
data fitted a series of five generalized linear mod-
els (GLMs): mainland Colombia as a whole;
Andes, Amazon, and Orinoco basins; and the
Caribbean region (Armenteras et al. 2013a,
Rodr
ıguez et al. 2013a). Here, we built an initial
GLM including data for the entire country and
the 15 explanatory variables previously analyzed
(Table 2). The percentage of initial forest cover in
each municipality was added as a covariate to
the previously published data, as research sug-
gests that it is a strong and negative covariate of
deforestation (Hargrave and Kis-Katos 2013).
We used four approaches to model deforesta-
tion as a function of covariates (Table 3). (1)
Generalized linear models were previously used
to analyze individual regions (Armenteras et al.
2013a, Rodr
ıguez et al. 2013a), and served as the
baseline lacking any method of accounting
intercepts and slope models (RIS) were used to
model deforestation on a regional level, but with-
out an explicit spatial component. To explicitly
account for spatial autocorrelation, we applied
(3) linear mixed-effects or hierarchical models
with correlation structures (geospatial LMEs);
and (4) Bayesian models with conditional autore-
gressive priors (CARs). The LME model accounts
for spatial autocorrelation using distance-based
correlation matrices within groups of observa-
tions, and the CAR priors use a binary neighbor-
based matrix. All models were implemented
using the open source R statistical language ver-
sion 3.0.1 (R Development Core Team 2008).
To reduce the number of explanatory variables
in the most computationally demanding of the
analyses accounting for spatial autocorrelation, an
initial Bayesian CAR analysis was conducted
using the CARBayes package (v4.1; Lee 2013). All

❖ www.esajournals.org 3 May 2017 ❖ Volume 8(5) ❖ Article e01824

 METS ET AL.

predictive variables, while the variables with

credible intervals including 0 were designated as
non-predictive. This reduced the potential predic-
tors to seven variables: elevation, legally protected
area, fire hotspots, urban population density,
unsatisfied basic needs (necesidades basicas insatis-
fechas [NBI]), road density, and initial percentage
of forest cover. These variables were then
included in subsequent RIS, LME, and CAR anal-
yses as predictors.

Spatial autocorrelation models

The RIS model is an extension of the GLM that
captures regional variation without explicitly
accounting for spatial autocorrelation. In the RIS
model, the intercept and slopes of different
covariates vary in the socio-environmental
regions of Colombia. This model was applied in
a Bayesian framework using JAGS version 3.4
(Plummer 2013) and implemented in R using the
R2jags package (v 0.5-6; Su and Yajima 2015).
The LME model included both a grouping fac-
tor by socio-environmental region and a geospa-
tial correlation matrix that applies within regions.
Fig. 1. Socio-environmental regions of Colombia. The geospatial correlation matrix of each LME
model assigns correlation weights to pairs of
potential explanatory variables were used as pre- observations determined by the distance between
dictors. The 95% high probability density (HPD) geographic centroids of the corresponding munic-
of coefficients was examined. Variables with HPD ipalities. The symmetric n 9 n correlation matrix
entirely below or above 0 were considered is set by a corSpatial object class from the nlme
package (v 3.1-109; Pinheiro et al. 2011). The corS-
Table 2. Coefficients and their P-values are given for patial object can use one of five different spatial
an initial GLM using 15 variables as covariates of functions to weigh the degree of spatial autocorre-
deforestation for the entire country. lation among residuals (Appendix S1: Fig. S1). We
Variables Coefficient P constructed LME models using each spatial decay
function with parameters estimated by maximum
Intercept 3.748 0.003
likelihood. The associated log likelihood of each
ln crops 0.057 0.070
ln pasture 0.089 0.231
LME model allowed for Akaike’s information
NBI 3.859 <0.001 criterion comparison with the exponential decay
Road density 0.342 <0.001 corExp model demonstrating best goodness of fit
Slope 0.012 0.784 (Appendix S1: Table S1).
Mining 0.004 0.234 The LME models allowed for estimates of coef-
Elevation 0.728 <0.001 ficients across all observations (equivalent to those
Precipitation <0.001 0.003
fitted in classical regressions and hereafter called
Water scarcity 0.007 0.959
Illicit crops 0.204 0.114
nationwide), alongside coefficients for individual
Fires 0.062 0.954 groups, hereafter called group-specific (Gelman
ln protected area 0.147 <0.001 2005, Gelman and Hill 2006). Two variables, forest
ln urban population 0.336 <0.001 cover and fire hotspots, had coefficients that chan-
ln rural density 0.252 0.051 ged signs among the different natural regions. For
ln cattle 0.027 0.827 the final LME model encompassing all seven pre-
Note: Bold values indicate P < 0.05. dictive variables, coefficients for forest cover and

❖ www.esajournals.org 4 May 2017 ❖ Volume 8(5) ❖ Article e01824


Table 3. Summary characteristics of the four models tested.
Characteristics GLM RIS
Implementation glm JAGS
Parameters 15 7 7
Grouping No Yes (7 parameters)
Autocorrelation No No
Note: GLM, generalized linear model; RIS, random intercepts and slopes; LME, linear mixed-effects model; CAR, condi-
tional autoregressive priors.
fires were estimated as specific to regions. The
remaining five predictive variables were estimated
as nationwide so as to reduce the computational
complexity of the LME model when applying the
correlation matrix. In every case, the geospatial
correlation matrix was applied to residuals of
observations within groups.
The second spatially explicit approach used a
Bayesian hierarchical model with CAR priors con-
structed with the CARBayes package (v 4.1; Lee
2013). A symmetric n 9 n neighborhood matrix
W sets the spatial autocorrelation structure of the
CAR priors for N observations. For observations
from neighboring spatial units k and i, wki = 1. If
the spatial units, municipalities in this study, do
not share a border, then wki = 0. The CAR priors
used for this study, originally modeled by Leroux
et al. (2000), are given by Eq. 2:
/k j/k ; W; s2 ; q
 P 
q n wki /i s2
 N Pn i¼1 ; Pn
q i¼1 wki þ 1  q q i¼1 wki þ 1  q
in which q is a correlation strength parameter
ranging from 0 to 1, and s is a parameter denot-
ing variance in the correlation weight. This
method does not take geographic distance
between municipalities into account. Coefficients
do not differ between the natural regions because
the data are not separated by regional groupings.
Comparing models
We compared the performance of the models
by (1) examining coefficients and their signifi-
cance or HPD, as spatial autocorrelation is
expected to inflate effect sizes for autocorrelated
variables and increase type I error rates; (2) esti-
mating spatial autocorrelation of residuals from
each model to diagnose incorrectly modeled spa-
tial variation; and (3) conducting posterior pre-
dictive checks of the models to evaluate model
❖ www.esajournals.org
METS ET AL.
LME with Geospatial matrix CAR
nlme, corStruct CARBayes
7
Yes (2 parameters) No
Yes (distance) Yes (neighbors)
precision. Spatial autocorrelation is expected to
increase homogeneity in observed values between
contiguous municipalities, especially when dis-
tances are small (Ord and Getis 1995), as in the
developed central Andes. To estimate residual
spatial autocorrelation, we used the ncf package
(v. 1.1-3; Bjørnstad 2009) to generate correlograms
for each model. The correlograms compared the
correlation of pairs of response estimates from
municipalities within designated distance inter-
vals from each other. For each 10-km distance
interval, 500 valid municipality pairs were ran-
domly sampled. The similarity in municipality
pairs among each distance interval was compared
to the nationwide similarity in deforestation val-
ues. Mean correlations significantly above the
nationwide baseline at low-distance intervals
indicated the presence of spatial autocorrelation.
Posterior predictive checks were completed for
the GLM, RIS, and CAR models implemented in
Bayesian analyses. The predictive checks used
coefficients sampled from the posterior distribu-
tions of each model to predict the deforestation
response variable. Using 1000 samples for each
municipality, we compared the predicted range to
the observation.
RESULTS
Covariates of change in forest cover
Statistically significant explanatory variables
in the all-variable GLM included demographic,
land-use, and geographic attributes (Table 1).
The variables unsatisfied basic needs, urban pop-
ulation density, and road density were all signifi-
cantly associated with deforestation. Fires,
protected area, and elevation were associated
with forest growth, while precipitation had sta-
tistically significant but minor effects on forest
growth.
5 May 2017 ❖ Volume 8(5) ❖ Article e01824

The best-fit geospatial LME model fitted an
exponential decay function to the spatial autocor-
relation of residuals. All variables with nationwide
coefficients had significant effects on change in
forest cover (P < 0.05). The covariates of deforesta-
tion were the same as for the GLM (Table 4). Pro-
tected area and elevation were similarly associated
with forest growth. Initial forest cover was mod-
eled as a region-specific predictor and changed
signs among the five regions (Table 4). In the
Amazon, Caribbean, and Pacific, initial forest
cover was a covariate of forest growth. In the
Andes and Orinoco, initial forest cover correlated
with deforestation. Fires was also modeled as
region-specific and were consistently opposite in
sign to the effect of initial forest cover, albeit to dif-
ferent extents in the five natural regions (Table 4).
Fires and initial forest cover had an overall corre-
lation value of 0.77 for the geospatial LME
model. Fires had a negative effect on deforestation
in the Andes and Orinoco basins. This result may
be associated with management and harvesting of
agriculture in these areas in addition to forest fires
propagated along the colonization front (Armen-
teras et al. 2011a).
The median coefficients of the CAR model
matched corresponding nationwide coefficients
of the LME model in sign and relative magnitude
(Table 4). Fires and forest cover were both linked
to forest growth. When compared to these spatial
models, the GLM estimated much greater effect
sizes for unsatisfied basic needs and urban popu-
lation density on deforestation.
The coefficients of the RIS model largely
agreed with those of the geospatial LME model.
Although there were slight differences in the
regional coefficients for land protection, eleva-
tion, unsatisfied basic needs, urban density, and
road density, there was no disagreement in sign.
The only notable difference was the coefficient
for fires in the Pacific region. Whereas fire was
associated with deforestation in the geospatial
LME model, it was associated with forest growth
in the RIS model.
Residual spatial autocorrelation
We calculated the correlations between the
residuals of pairs of municipalities at bins of geo-
graphic distance between municipal centroids to
obtain correlograms (Fig. 2). Pairs of municipalities
❖ www.esajournals.org
METS ET AL.
in close proximity to each other had an overall
significantly positive correlation of residuals in
the GLM. This effect persisted for municipality
pairs separated by up to 50 km. Municipality
pairs in the RIS model shared the general trend
of the GLM, with significant correlation of resid-
uals in close proximity, albeit to a reduced extent.
The geospatial LME and CAR models displayed
different trends, and there was no significant
correlation of residuals in pairs of municipalities
in close proximity.
Bayesian posterior predictive checks
The CAR and RIS approaches were Bayesian
models and thus had posterior distributions for
each predictive coefficient. A Bayesian implemen-
tation of the GLM had median coefficients
(Appendix S1: Table S2) identical to those esti-
mated using the glm function in R. Coefficients
were sampled from the posterior distributions of
each model and used to predict the rate of defor-
estation of each municipality. These predictions
were compared to the observed rates of deforesta-
tion (Fig. 3). The geospatial LME model required
computation of an n 9 n matrix of distance-based
correlation weights for the 1119 municipalities,
which presented tremendous computational com-
plexity with a Bayesian implementation. Hence
the geospatial LME model was excluded from this
analysis.
The models displayed similar trends in predic-
tions for the Andes, Amazon, and Orinoco regions
(Fig. 3). The full GLM consistently underesti-
mated deforestation rates and excluded approxi-
mately half of all real observations from the
predictive ranges. The RIS model similarly under-
estimated the rate of deforestation. The general
trend of underestimating deforestation for both
these models held for predictions across all
regions. The CAR model predictions, while shar-
ing the tendency to underestimate deforestation,
encapsulated most real observations within pre-
dictive ranges. The extreme ends of the Andes, for
municipalities that experienced the most defor-
estation or forest growth, fell well outside any pre-
dictive posteriors. The Caribbean region displayed
the most deviation between predicted and
observed deforestation rates. All approaches failed
to predict deforestation observed in three-quarters
of the municipalities in the Caribbean region.
6 May 2017 ❖ Volume 8(5) ❖ Article e01824
 METS ET AL.

Table 4. Coefficients for intercept and seven covariates.

Covariates GLM CAR LME with Geospatial matrix RIS

Intercept 3.041 1.622 (0.851, 4.155)

Ama. 1.709 1.790 (1.035, 4.635)
And. 0.413 1.479 (1.357, 4.139)
Car. 3.210 2.103 (0.676, 5.071)
Ori. 0.048 1.698 (1.318, 4.570)
Pac. 2.836 1.615 (1.291, 4.387)
Forest cover 0.457* 1.166 (2.309, 0.040)
Ama. 1.151 0.384 (3.050, 1.701)
And. 0.977 1.012 (0.305, 2.711)
Car. 3.655 1.447 (5.079, 1.094)
Ori. 1.609 0.337 (2.255, 3.181)
Pac. 3.015 0.355 (2.392, 1.386)
Fires 0.623 1.652 (3.795, 0.520)
Ama. 0.800 1.794 (0.312, 5.188)
And. 0.915 4.083 (7.792, 0.507)
Car. 2.408 1.307 (2.286, 4.661)
Ori. 1.334 4.498 (9.236, 0.212)
Pac. 2.139 3.403 (10.694, 2.731)
Protection 0.131** 0.161 (0.231, 0.091) 0.155**
Ama. 0.193 (0.045, 0.060)
And. 0.072 (0.166, 0.021)
Car. 0.336 (0.502, 0.172)
Ori. 0.146 (0.353, 0.073)
Pac. 0.194 (0.324, 0.067)
Elevation 0.771** 0.191 (0.457, 0.071) 0.290**
Ama. 0.387 (0.695, 0.012)
And. 0.395 (0.636, 0.131)
Car. 0.468 (0.769, 0.202)
Ori. 0.364 (0.671, 0.059)
Pac. 0.418 (0.687, 0.133)
NBI 3.674** 1.451 (0.093, 2.816) 2.040**
Ama. 1.899 (0.376, 3.732)
And. 1.923 (0.335, 3.399)
Car. 2.914 (0.980, 5.848)
Ori. 1.859 (0.578, 3.869)
Pac. 2.246 (0.649, 3.951)
Urban density 0.268** 0.081 (0.044, 0.208) 0.112
Ama. 0.133 (0.033, 0.303)
And. 0.108 (0.035, 0.250)
Car. 0.186 (0.007, 0.411)
Ori. 0.130 (0.090, 0.358)
Pac. 0.061 (0.155, 0.247)
Road density 0.348** 0.257 (0.086, 0.427) 0.362**
Ama. 0.532 (0.523, 1.513)
And. 0.261 (0.081, 0.439)
Car. 0.547 (0.222, 0.867)
Ori. 0.288 (0.474, 0.942)
Pac. 1.179 (0.751, 1.604)
Notes: GLM, generalized linear model; RIS, random intercepts and slopes; LME, linear mixed-effects model; CAR,
conditional autoregressive priors. Ama, Amazon; And, Andes; Car, Caribbean; Ori, Orinoco; Pac, Pacific. Median coefficient
values are displayed for the RIS and CAR models. Positive values indicate variables associated with deforestation, and negative
values indicate variables associated with forest growth. Median coefficients shown for CAR and RIS with 95% of the highest
probability density shown in parentheses.

P < 0.05,  P < 0.01 where applicable to nationwide coefficients of GLM and LME.

❖ www.esajournals.org 7 May 2017 ❖ Volume 8(5) ❖ Article e01824


Fig. 2. Correlograms of residual correlation among municipality pairs separated by distance classes defined
by km distance between paired municipality centroids. Closed dots indicate average correlation values signifi-
cantly above the national baseline.
Heterogeneity of deforestation covariates in the
Caribbean region
Deviance between observations and the predic-
tive ranges was largest in the Caribbean region
for all three models tested using posterior predic-
tive checks. The geographic distribution of forests
in the Caribbean region centers in two clusters:
the protected Sierra Nevada de Santa Marta in the
north and the Serran
ıa de San Lucas at the north-
ernmost end of the central Andes. To test whether
deforestation rates in these two clusters have
different covariates, we divided the Caribbean
region along the Magdalena River into western
(containing the Santa Marta area) and eastern
(containing the San Lucas area) blocs. A Bayesian
❖ www.esajournals.org
METS ET AL.
model with group-specific intercepts and slopes
was then constructed to predict the rate of forest
change in the Caribbean municipalities based on
the same set of predictive parameters as in the
other nationwide models.
The posterior distributions for the group-
specific coefficients were compared between the
eastern and western portions of the Caribbean
natural region. While posterior distributions for
the model intercept and most covariates overlay
each other (Fig. 4), road density had a stronger
association with deforestation in the eastern
Caribbean, while fires was considerably more
associated with deforestation in the western
Caribbean.
8 May 2017 ❖ Volume 8(5) ❖ Article e01824
 METS ET AL.

Fig. 3. Posterior predictive checks of three models. For each municipality and model, coefficients are sampled
from posterior distributions and used to predict the rate of change in forest cover. Dark points represent observed
rates of deforestation for each municipality ordered from least to greatest deforestation. Lighter points represent
1000 predicted rates of deforestation for each municipality as estimated through sampled coefficient values.

DISCUSSION models, and (2) implications for recent analyses

Our analyses highlight the impact of modeling
choices on model coefficients and their statistical
significance, as well as the explanatory power of
models. As geographic analyses inform concrete
actions (Nepstad et al. 2006, Soares-Filho et al.
2010, Aide et al. 2013, Nolte et al. 2013), method-
ological choices can have profound consequences
in shaping international initiatives such as
REDD+, or national policies such as designation
of land for legal protection (Portocarrero-Aya
et al. 2014). Capturing the geographic structure
of observations is thus an important requirement
to model deforestation for policy decisions. We
focus on two key findings: (1) change in model
performance, in particular the systemic underes-
timation of deforestation rates from non-spatial
of deforestation in Colombia.
Model performance
Posterior predictive checks reveal that autocor-
relation not captured by models biases predictions
(Fig. 3), and prediction improves with spatially
explicit methods such as CAR. But there are many
methods to quantify and address residual autocor-
relation (Cliff and Ord 1970, Lam 1983, Portocar-
rero-Aya et al. 2014), and how to interpret results
from different approaches is often uncertain (Getis
2007). A comparison between the LME geospatial
and the CAR models illustrates the difference
between models. The CAR model did not separate
the municipalities by natural region, so compar-
isons are possible for the nationwide coefficients,
but not for the region-specific parameters of the

❖ www.esajournals.org 9 May 2017 ❖ Volume 8(5) ❖ Article e01824


Fig. 4. Density plots of four coefficients differing between the eastern (solid line) and western (dotted line) Car-
ibbean as separated by the Magdalena River. The vertical dotted line indicates the median coefficient value when
the Caribbean is modeled as a single region.
LME. Despite using fundamentally different mea-
sures of spatial autocorrelation, the two models
estimated similar coefficients for the nationwide
predictive variables (Table 4). That is, there was
no great impact on modeling the effects of covari-
ates whether the correlation matrix was based on
distance or on sharing of borders.
Spatial autocorrelation between observations is
inversely correlated with distance and thus would
be most prevalent in the densely populated clus-
ter of the Andes where distances between munici-
palities are shortest. This central part of the
country is where the Colombian state is most
effective and data were anticipated to be most
reliable (Mainwaring 2006). In addition, the
Andes natural region contains the majority of
municipalities, 631 out of 1119. This introduces a
twofold problem for nationwide analyses: Model
results may be biased toward deforestation deter-
minants unique to the largely settled Andean
region, and the spatial clustering of the Andes
municipalities may inflate coefficients failing to
describe deforestation in the more sparsely popu-
lated and heavily forested natural regions.
❖ www.esajournals.org
METS ET AL.
The effects of spatial autocorrelation with clus-
tered units such as the Andes are best illustrated
by comparing the CAR and RIS models. The
CAR model outperforms the RIS model in all
natural regions, despite the RIS model estimating
region-specific coefficients (Fig. 3). Therefore,
modeling the effect of spatial autocorrelation in
the Andes improves precision for the outlying
natural regions as well. The most predictive
covariates of deforestation, NBI and road density,
proved unidirectional in the RIS model and
maintained similar magnitude across the differ-
ent natural regions (Table 4). Because of the con-
sistency of these covariates, the CAR model did
not lose predictive power in outlying regions
compared to the RIS, despite estimating only
nationwide coefficients.
The optimal corStruct models applied for the
LME spatial correlation matrix (Appendix S1:
Fig. S1) show that spatial autocorrelation of residu-
als from deforestation data dissipates rapidly with
distance. The best-performing model, an exponen-
tial decay function, had the correlation weight of
municipality pairs approaching the national
10 May 2017 ❖ Volume 8(5) ❖ Article e01824

baseline within just 30 km distance. In developing
regions of Colombia, municipality centroids are
typically separated from neighboring centroids by
well over 30 km. Observations from those regions
where municipalities are larger will not exhibit
considerable within-region spatial autocorrelation.
This is problematic for the CAR model, which
ignores distance in lieu of municipality neighbor-
hood. However, while CAR-determined covariates
of deforestation are reduced in magnitude com-
pared to LME coefficients, there is agreement in
sign and relative magnitude between variables
(Table 4). This suggests that the CAR model,
although ignoring physical distance between
observations, does not overcompensate in applying
correlation weights in regions with large munici-
palities such as in the Amazon and Orinoco basins.
In short, while LME models have the advan-
tage of estimating group-specific coefficients, the
absence of this structure does not undermine
inference of whole-sample coefficients in CAR
models. Both are preferable to non-spatial mod-
els for estimating the statistical relevance and
effect size of predictors, and improving model
precision. Failing to account for spatial autocor-
relation in analyses of deforestation across units
with skewed size distributions (as is common in
analyses of political units) risks estimating
parameters based on clusters of similar neighbor-
ing units. Finally, we demonstrated the use of
posterior predictive checks to determine model
precision as a powerful approach for uncovering
region-specific misspecification and improve pre-
diction of deforestation rates.
National determinants of deforestation
All four models estimated similar trends for the
five nationwide variables of the LME model:
elevation, land protection, urban population den-
sity, unsatisfied basic needs, and road density.
Unsatisfied basic needs, used in large part as a
proxy for areas at the forefront of colonization,
was the variable most strongly correlated with
deforestation. This index measures the percentage
of the municipal population that lacks access to
sanitary services, primary education, and mini-
mum household economic capacity. Municipali-
ties with high NBI largely lack state presence and
often correspond to the rapidly changing agricul-
tural frontier between settled and relatively
developed cores, and newly colonized areas of
❖ www.esajournals.org
METS ET AL.
primarily indigenous occupation (Rudel and
Roper 1997, Rodr
ıguez et al. 2012). The coefficient
of NBI indicates that increasing unsatisfied basic
needs promote deforestation. There is a wealth of
literature connecting the agricultural frontier to
land-use change (Scherr 2000, Barbier 2012a, Poko-
rny et al. 2013). Briefly, roads and waterways
serve as means for colonist migration sometimes
promoted through specific projects (e.g., oil devel-
opment, or allocation of land titles) to abundant
forested land (Rudel and Roper 1997). Newly
settled areas are then quickly cleared to establish
ownership, extract as much of its natural
resources as possible, or both (Southgate 1990,
Fearnside 2005). Both population growth and pov-
erty contribute to deforesting this frontier, not as
driving factors, but as sources of smallholders to
colonize newly opened lands (Fearnside 1993,
Lambin et al. 2001). The end result of frontier set-
tlement is a largely deforested landscape
(Rodr
ıguez et al. 2012), sometimes economically
developed but more often not (Barbier 2012b).
Poor financial returns from smallholder agricul-
ture, policies that effectively promote consolidated
landholdings and ranching, and more forested
land available discourage long-term sustainable
practices and promote further deforestation
(Hecht 1993, Coomes et al. 2011). The strong asso-
ciation between NBI and deforestation supports
policy initiatives aiming to shift incentives against
forest exploitation, and promote sustainable land
ownership and practices at the frontier (Rodrigues
et al. 2009, Dulal et al. 2012).
Additionally, in Colombia, the forest frontier
and high NBI are also associated with the pres-
ence of armed groups. Armed conflict resulting in
population displacement concentrates poverty
into areas where state institutions are least effec-
tive (D
ıaz and S
anchez 2004, Ib
an ~ ez and Moya
2010). Municipalities with reduced state presence
are also more likely to have forest cleared for coca
cultivation (Dion and Russler 2008). Our results
did not find illicit crops to be a significant predic-
tor of deforestation (Table 2). Multiple analyses
with independent deforestation datasets have
found that coca cultivation does not explain
deforestation once socioeconomic characteristics
are included as covariates (D
avalos et al. 2011,
S
anchez-Cuervo and Aide 2013). Violent conflict
and displacement partly reflected in NBI and
otherwise absent from our models are important
11 May 2017 ❖ Volume 8(5) ❖ Article e01824

predictors of deforestation in Colombia (Sanchez-
Cuervo and Aide 2013, Fergusson et al. 2014).
These findings suggest that communities in
the process of development exhibit the highest
rates of deforestation (Rodr
ıguez et al. 2012).
Continued colonization of forested land, whether
from exploitation of land resources and/or dis-
placement caused by political instability (Lo
pez-
Carr 2008), will maintain rates of deforestation
despite increasing urbanization. The rapidly
changing and receding forest frontier of rural col-
onization is shared across many developing trop-
ical countries with abundant forests (Rudel and
Roper 1997, Barbier 2004, Lo
pez-Carr and
Burgdorfer 2013). With surrogates for coloniza-
tion as the most potent driver of deforestation in
our analyses, initiatives to curtail deforestation
must stabilize the advancing frontier and pro-
vide economic incentives for conservation among
rural communities (Blom et al. 2010).
The results demonstrate that urban population
density and road density are significantly associ-
ated with deforestation. The density of road net-
works, which allows for easier resource
extraction and encourages the conversion of for-
ests to pastures (De Luca 2007, Barber et al. 2014,
Newman et al. 2014), is consistently associated
with deforestation, but to a lesser extent than
NBI. Road development, both paved and
unpaved, has been repeatedly identified as a key
step to the clearing of forests (Pfaff 1999, Soares-
Filho et al. 2004), providing access into otherwise
impenetrable forest regions in less developed
municipalities (Arima et al. 2005, Pfaff et al.
2007). Forest patches closer to developed urban
areas, particularly developing regions, such as
the Amazon, are more convenient targets to ille-
gal deforestation (Laurance et al. 2002). Settle-
ment of areas beyond the more developed
Andean region required initial spontaneous or
directed colonization followed by local resource
extraction (Schuurman 1978, Rudel 2007). Urban-
ization represents a later stage in settlement
when demographic cores are more fully estab-
lished and growing. At this point, local land
value increases as economic incentives pull
toward converting forest to pasture and indus-
trial agriculture (Rudel et al. 2009, Seto et al.
2010, D
avalos et al. 2014).
The proportion of area protected had a nega-
tive association with deforestation. Although
❖ www.esajournals.org
METS ET AL.
protected areas are not necessarily free from
deforestation (Armenteras et al. 2011b, Rodr
ıguez
et al. 2013b), our results confirm nationally pro-
tected land, including indigenous reserves, fare
better against deforestation than territories with-
out legal protection. This has been observed in the
Amazon region of Brazil as well (Nepstad et al.
2006). Globally, the establishment of protected
areas reduces deforestation although deterrent
effects tend to be weaker in areas further from
roads and urban areas (Joppa and Pfaff 2010,
Geldmann et al. 2013). The association between
forest growth and protected areas is strongest in
the Caribbean and Pacific regions (Table 4), which
are slightly more developed than in the Orinoco
and Amazon basins. Legal protection impedes the
development of roads, both paved and unpaved,
which are critical for supporting deforestation
operations (Nepstad et al. 2009). Additionally,
insecure property rights promotes deforestation
by providing incentive for landholders to clear
forests, grow crops, and build structures to claim
land (Araujo et al. 2009). Legal protection pre-
vents this ambiguity in property rights (Fearnside
2001). Analyses based in Costa Rica and Thailand
have indicated that protected areas promote
reduction in local poverty (Andam et al. 2010).
The mechanism for this alleviation in poverty is
increased tourism (Ferraro and Hanauer 2014),
which enforces economic incentives for keeping
forests intact.
Increased deforestation rates of lowlands likely
drive the negative association between elevation
and deforestation. Colombia’s colonial-era settle-
ment started along the Andes mountain range
and along the Caribbean shore, so that natural
forests were already reduced in the densely pop-
ulated mid-elevation regions of the country by
the beginning of the 20th century (Etter et al.
2008). Exploitation of remnant lowland forests
has increased on a global scale in the past few
decades with rising urban populations and
increased international trade (DeFries et al.
2010). Present major targets of deforestation in
the Pacific and the Amazon regions are in low-
elevation areas (Etter et al. 2006). Additionally,
areas of reforestation have been identified in
high-elevation parts of the Andes (S
anchez-
Cuervo et al. 2012, Sanchez-Cuervo and Aide
2013). Our results corroborate these findings
across all natural regions.
12 May 2017 ❖ Volume 8(5) ❖ Article e01824

Regional determinants of deforestation
Forest cover and fires were modeled as region-
specific variables for the geospatial LME model
because their sign changed between natural
regions in the RIS model. When region-specific
effects were not modeled, initial forest cover and
fires held a positive association with forest
growth (Table 4). Initial forest cover is a measure
of how much forest cover was available for
exploitation, and fires indirectly measures clear-
cutting activity, though the proportion of fires
associated with forest clearing may differ among
Colombia’s natural regions (Armenteras et al.
2011a). Deforestation builds upon prior activity
in the more well-developed sections of the natu-
ral regions (Soares-Filho et al. 2004), and our
analyses were consistent with this pattern. In the
spatially explicit LME model, municipalities with
greater proportions of forest cover experienced
reduced deforestation in the Amazon, Caribbean,
and Pacific natural regions. The Amazon and
Pacific regions had larger proportions of munici-
palities with forest cover at the start of the study
period (Appendix S1: Table S3). In these lowland
regions, municipalities with little forest cover are
part of the agricultural frontier and experience
rapid land-use change. The effect of the forested
proportion was the opposite in the Andes and
Orinoco basins, both of which generally have
low forest cover (Appendix S1: Fig. S2). The
Andes are heavily developed and forests have
been reduced to smaller remnants decades ago
(Etter and Villa 2000, Armenteras et al. 2011b).
The Orinoco Basin has wide swaths of natural
savanna interspersed with gallery forests that
make up a smaller proportion of the municipali-
ties there. The deforestation that does occur in
both of these regions is located in municipalities
with large remaining patches of forest. These
results highlight the importance of exploring
region-specific effects, as whole-sample analyses
of municipalities (e.g., D

avalos et al. 2011) would
fail to model these important differences.
Fires exhibited sign change between natural
regions with the RIS model. The signs of the coef-
ficient for fires between the RIS and LME models
remained the same with the exception of the
Pacific natural region (Table 4). The Pacific coast,
one of the wettest regions in the world with
mean annual precipitation ranging from 3000 to
12,000 mm, records few fires, and thus, the
❖ www.esajournals.org
METS ET AL.
model coefficient for fires had little effect on the
prediction of deforestation rates. The strongest
signal for fires as a covariate of deforestation
comes from the western portion of the Caribbean
natural region (Fig. 4). Fires in the western Car-
ibbean region are associated with clear-cutting
preceding development of agricultural land, and
this association with deforestation is lost for
municipalities east of the Magdalena River.
Clear-cutting in the western Caribbean coast is a
destructive yet locally constrained phenomenon
(Chadid et al. 2015). The effect of fires is not as
strong elsewhere in the country. This makes the
western Caribbean an ideal region for using
MODIS remote sensing data to detect forest fires
and identify sites of potentially ongoing defor-
estation (Yu et al. 2005). Prior use of moderate-
resolution imaging spectroradiometer, or MODIS
data to understand associations between fires
and deforestation in the Colombian Amazon has
demonstrated the use of fires for conversion of
forest to pastures (Armenteras et al. 2013b).
While other variables maintained sign consis-
tency in relation to deforestation, the magnitude
of coefficient values exhibited differences between
natural regions. This is particularly true for the
coefficients of protected areas and road density,
for which the median values of certain natural
regions fell outside the highest probability density
interval of other regions (Table 4). This indicates
important regional differences in the effects of
these variables. Important related attributes that
have not been included in the models, such as
enforcement of protected areas (Pfaff et al. 2014),
and differentiation between settlement types for
road density (Perz et al. 2013), may differ between
the natural regions and cause the deviation seen
in median coefficient values.
CONCLUSIONS
We compared four models of deforestation dif-
fering in their treatment of spatial autocorrelation.
Analyses of autocorrelation show that non-spatial
models will incur deflated residual variance, and
are thus subject to high type I error for spatially
clustered covariates. The failure of RIS models
indicates that modeling covariates at a region-
specific scale is not enough to correct for the
underlying spatial structure of the data, resulting
in model bias. Using geospatial LME or CAR
13 May 2017 ❖ Volume 8(5) ❖ Article e01824

models effectively addresses these biases. Poste-
rior predictive checks demonstrate that the spa-
tially explicit CAR model outperforms non-spatial
GLM and RIS models.
Compared to previous analyses of deforestation
in Colombia, our analyses revealed that differ-
ences in coefficients for deforestation covariates
do not always correspond to large natural
regions. Critical differences between the eastern
and western Caribbean regions were identified
using posterior predictive checks. Differences in
coefficient values were also found between natu-
ral regions for protected areas and road density,
indicating that these variables may be further
affected by unmodeled landscape features. While
all models failed to predict outliers of extreme
deforestation or relatively high forest growth,
these spatially explicit analyses further link fron-
tier development and road construction to defor-
estation. Additionally, legal protection of land has
a demonstrable positive association with forest
growth. Continued protection, and further exten-
sion to high-risk areas, is a critical component of
forest conservation in Colombia.
ACKNOWLEDGMENTS
We thank L. Yohe, M. Lim, and E. Lauterbur for
helpful comments on the manuscript. We also thank
the Instituto de Hidrolog
ıa, Meterorolog
ıa y Estudios
Ambientales (IDEAM), for providing data maps.
LITERATURE CITED
Aide, T. M., M. L. Clark, H. R. Grau, D. Lo
pez-Carr,
M. A. Levy, D. Redo, M. Bonilla-Moheno, G. Riner,
M. J. Andrade-Nu
n
~ ez, and M. Mu~ niz. 2013. Defor-
estation and reforestation of Latin America and the
Caribbean (2001–2010). Biotropica 45:262–271.
Andam, K. S., P. J. Ferraro, K. R. Sims, A. Healy, and
M. B. Holland. 2010. Protected areas reduced pov-
erty in Costa Rica and Thailand. Proceedings of the
National Academy of Sciences of the United States
of America 107:9996–10001.
Araujo, C., C. A. Bonjean, J.-L. Combes, P. C. Motel,
and E. J. Reis. 2009. Property rights and deforesta-
tion in the Brazilian Amazon. Ecological Eco-
nomics 68:2461–2468.
Arima, E. Y., R. T. Walker, S. G. Perz, and M. Caldas.
2005. Loggers and forest fragmentation: behavioral
models of road building in the Amazon basin.
Annals of the Association of American Geogra-
phers 95:525–541.
❖ www.esajournals.org
METS ET AL.
Armenteras, D., E. Cabrera, N. Rodr
ıguez, and
J. Retana. 2013a. National and regional determi-
nants of tropical deforestation in Colombia. Regio-
nal Environmental Change 13:1181–1193.
Armenteras, D., J. Retana-Alumbreros, R. Molowny-
Horas, R. M. Roman-Cuesta, F. Gonzalez-Alonso,
and M. Morales-Rivas. 2011a. Characterising fire
spatial pattern interactions with climate and vege-
tation in Colombia. Agricultural and Forest Meteo-
rology 151:279–289.
Armenteras, D., N. Rodr
ıguez, and J. Retana. 2013b.
Landscape dynamics in northwestern Amazonia:
an assessment of pastures, fire and illicit crops as
drivers of tropical deforestation. PLoS ONE 8:
e54310.
Armenteras, D., N. Rodr
ıguez, J. Retana, and
M. Morales. 2011b. Understanding deforestation in
montane and lowland forests of the Colombian
Andes. Regional Environmental Change 11:693–705.
Barber, C. P., M. A. Cochrane, C. M. Souza, and W. F.
Laurance. 2014. Roads, deforestation, and the
mitigating effect of protected areas in the Amazon.
Biological Conservation 177:203–209.
Barbier, E. B. 2004. Explaining agricultural land expan-
sion and deforestation in developing countries.
American Journal of Agricultural Economics 86:
1347–1353.
Barbier, E. B. 2012a. Natural capital, ecological scarcity
and rural poverty. World Bank Policy Research
Working Paper. https://doi.org/10.1596/1813-9450-
6232
Barbier, E. B. 2012b. Scarcity, frontiers and develop-
ment. Geographical Journal 178:110–122.
Beale, C. M., J. J. Lennon, J. M. Yearsley, M. J. Brewer,
and D. A. Elston. 2010. Regression analysis of spa-
tial data. Ecology Letters 13:246–264.
Bjørnstad, O. 2009. ncf: spatial nonparametric covari-
ance functions. R package v 1.1-3.
Blom, B., T. Sunderland, and D. Murdiyarso. 2010. Get-
ting REDD to work locally: lessons learned from
integrated conservation and development projects.
Environmental Science & Policy 13:164–172.
Brown, D. G., P. H. Verburg, R. G. Pontius, and M. D.
Lange. 2013. Opportunities to improve impact,
integration, and evaluation of land change models.
Current Opinion in Environmental Sustainability
5:452–457.
Chadid, M. A., L. M. D
avalos, J. Molina, and
D. Armenteras. 2015. A Bayesian spatial model
highlights distinct dynamics in deforestation from
coca and pastures in an Andean biodiversity
hotspot. Forests 6:3828–3846.
Cliff, A. D., and K. Ord. 1970. Spatial autocorrelation:
a review of existing and new measures with appli-
cations. Economic Geography 46:269–292.
14 May 2017 ❖ Volume 8(5) ❖ Article e01824

Cliff, A., and K. Ord. 1972. Testing for spatial autocor-
relation among regression residuals. Geographical
Analysis 4:267–284.
Coomes, O. T., Y. Takasaki, and J. M. Rhemtulla. 2011.
Land-use poverty traps identified in shifting cultiva-
tion systems shape long-term tropical forest cover.
Proceedings of the National Academy of Sciences of
the United States of America 108:13925–13930.
DANE [Departamento Administrativo Nacional de
Estad
ıstica]. 1985. Censo nacional. Departamento
Administrativo Nacional de Estad
ıstica, Bogot
a,
D.C., Colombia.
DANE [Departamento Administrativo Nacional de
Estad
ıstica]. 1993. Censo nacional de poblacio
n y
vivienda. Departamento Administrativo Nacional
de Estad
ıstica, Bogot
a, D.C., Colombia.
DANE [Departamento Administrativo Nacional de
Estad
ıstica]. 2008. Sistema de Consulta Informacio
n
Censal: Censo 2005. Departamento Administrativo
Nacional de Estad
ıstica, Bogot
a, D.C., Colombia.
D
avalos, L. M., A. C. Bejarano, M. A. Hall, H. L.
Correa, A. Corthals, and O. J. Espejo. 2011. Forests
and drugs: coca-driven deforestation in tropical
biodiversity hotspots. Environmental Science &
Technology 45:1219–1227.
D
avalos, L. M., J. S. Holmes, N. Rodr
ıguez, and
D. Armenteras. 2014. Demand for beef is unrelated
to pasture expansion in northwestern Amazonia.
Biological Conservation 170:64–73.
De Luca, G. D. 2007. Roads, development and defor-
estation: a review. World Bank Development
Research Group Paper, World Bank and CRED,
University of Namur, Washington, D.C., USA and
Namur, Wallonia, Belgium.
DeFries, R. S., T. Rudel, M. Uriarte, and M. Hansen.
2010. Deforestation driven by urban population
growth and agricultural trade in the twenty-first
century. Nature Geoscience 3:178–181.
D
ıaz, A. M., and F. S
anchez. 2004. A geography of illi-
cit crops (coca leaf) and armed conflict in Colom-
bia. Development Research Centre, London, UK.
Dion, M. L., and C. Russler. 2008. Eradication efforts,
the state, displacement and poverty: explaining
coca cultivation in Colombia during Plan Colom-
bia. Journal of Latin American Studies 40:399–421.
Dulal, H. B., K. U. Shah, and C. Sapkota. 2012. Reducing
emissions from deforestation and forest degradation
(REDD) projects: lessons for future policy design
and implementation. International Journal of Sus-
tainable Development & World Ecology 19:116–129.
Etter, A., C. McAlpine, and H. Possingham. 2008.
Historical patterns and drivers of landscape change
in Colombia since 1500: a regionalized spatial
approach. Annals of the Association of American
Geographers 98:2–23.
❖ www.esajournals.org
METS ET AL.
Etter, A., C. McAlpine, D. Pullar, and H. Possingham.
2006. Modelling the conversion of Colombian lowland
ecosystems since 1940: drivers, patterns and rates.
Journal of Environmental Management 79:74–87.
Etter, A., and L. A. Villa. 2000. Andean forests and
farming systems in part of the Eastern Cordillera
(Colombia). Mountain Research and Development
20:236–245.
Fearnside, P. M. 1993. Deforestation in Brazilian Ama-
zonia: the effect of population and land tenure.
Ambio-Journal of Human Environment Research
and Management 22:537–545.
Fearnside, P. M. 2001. Land-tenure issues as factors in
environmental destruction in Brazilian Amazonia:
the case of southern Par
a. World Development
29:1361–1372.
Fearnside, P. M. 2005. Deforestation in Brazilian
Amazonia: history, rates, and consequences.
Conservation Biology 19:680–688.
Fergusson, L., D. Romero, and J. F. Vargas. 2014.
The environmental impact of civil conflict: the
deforestation effect of paramilitary expansion in
Colombia. Serie Documento CEDE, 2014-36.
Centro de Estudios sobre Desarollo Econo
mico,
Bogot
a, Colombia.
Ferraro, P. J., and M. M. Hanauer. 2014. Quantifying
causal mechanisms to determine how protected
areas affect poverty through changes in ecosystem
services and infrastructure. Proceedings of the
National Academy of Sciences of the United States
of America 111:4332–4337.
Geldmann, J., M. Barnes, L. Coad, I. D. Craigie,
M. Hockings, and N. D. Burgess. 2013. Effective-
ness of terrestrial protected areas in reducing habitat
loss and population declines. Biological Conserva-
tion 161:230–238.
Gelman, A. 2005. Analysis of variance-why it is more
important than ever. Annals of Statistics 33:1–53.
Gelman, A., and J. Hill. 2006. Data analysis using
regression and multilevel/hierarchical models. Cam-
bridge University Press, New York, New York, USA.
Getis, A. 2007. Reflections on spatial autocorrelation.
Regional Science and Urban Economics 37:491–496.
Gullison, R. E., P. C. Frumhoff, J. G. Canadell, C. B.
Field, D. C. Nepstad, K. Hayhoe, R. Avissar, L. M.
Curran, P. Friedlingstein, and C. D. Jones. 2007.
Tropical forests and climate policy. Science 316:985.
Hargrave, J., and K. Kis-Katos. 2013. Economic causes
of deforestation in the Brazilian Amazon: a panel
data analysis for the 2000s. Environmental and
Resource Economics 54:471–494.
Hecht, S. B. 1993. The logic of livestock and deforesta-
tion in Amazonia. BioScience 43:687–695.
Hijmans, R. J., S. E. Cameron, J. L. Parra, P. G. Jones,
and A. Jarvis. 2005. Very high resolution interpolated
15 May 2017 ❖ Volume 8(5) ❖ Article e01824

climate surfaces for global land areas. International
Journal of Climatology 25:1965–1978.
Hunter, G. J., A. K. Bregt, G. B. Heuvelink, S. De Bruin,
and K. Virrantaus. 2009. Spatial data quality: prob-
lems and prospects. Pages 101–121 in G. Navratil,
editor. Research trends in geographic information
science. Springer, Berlin, Germany.
Ib
an
~ ez, A. M., and A. Moya. 2010. Do conflicts create
poverty traps? Asset losses and recovery for dis-
placed households in Colombia. Pages 137–172 in
R. Di Tella, S. Edwards, and E. Schargrodsky,
editors. The economics of crime: lessons for and
from Latin America. University of Chicago Press,
Chicago, Illinois, USA.
IDEAM. 2000. Informe nacional del agua. Instituto de
Estudios Ambientales y Meteorologicos, Bogot
a,
D.C., Colombia.
IDEAM, IGAC, IAvH, Invemar, Sinchi, IIAP. 2007. Eco-
sistemas continentales, costeros y marinos de
Colombia. Instituto de Hidrolog
ıa, Meteorolog
ıa y
Estudios Ambientales, Instituto Geogr
afico
Agust
ın Codazzi, Instituto de Investigacio
n de
Recursos Biolo
gicos Alexander von Humboldt,
Instituto de Investigaciones Marinas y Costeras
Jos
e Benito Vives de Andr
eis, Instituto Amazo
nico
de Investigaciones Cient
ıficas Sinchi e Instituto de
Investigaciones Ambientales del Pac
ıfico Jhon von
Neumann, Bogot
a, D.C., Colombia.
IGAC [Instituto Geogr
afico Agust
ın Codazzi]. 2005.
Cartograf
ıa b
asica escala 1:500.000 (cubrimiento
nacional). Instituto Geogr
afico Agust
ın Codazzi,
Bogot
a, D.C., Colombia.
IGAC [Instituto Geogr
afico Agust
ın Codazzi]. 2011.
Sistema de informacio
n geogr
afica para la
planeacio
n y el ordenamiento territorial nacional
SIGOT. Instituto Geogr
afico Agust
ın Codazzi,
Bogot
a, D.C., Colombia.
Joppa, L. N., and A. Pfaff. 2010. Global protected area
impacts. Proceedings of the Royal Society of Lon-
don. Series B: Biological Sciences 278:1633–1638.
Kaimowitz, D., B. Mertens, S. Wunder, and P. Pacheco.
2004. Hamburger connection fuels Amazon destruc-
tion. Center for International Forest Research,
Bangor, Indonesia.
Koenig, W. D. 1999. Spatial autocorrelation of ecologi-
cal phenomena. Trends in Ecology & Evolution
14:22–26.
Kruskal, W. 1988. Miracles and statistics: the casual
assumption of independence. Journal of the Ameri-
can Statistical Association 83:929–940.
Ku€ hn, I., and C. F. Dormann. 2012. Less than eight
(and a half) misconceptions of spatial analysis.
Journal of Biogeography 39:995–998.
Lam, N. S.-N. 1983. Spatial interpolation methods: a
review. American Cartographer 10:129–150.
❖ www.esajournals.org
METS ET AL.
Lambin, E. F., B. L. Turner, H. J. Geist, S. B. Agbola,
A. Angelsen, J. W. Bruce, O. T. Coomes, R. Dirzo,
G. Fischer, and C. Folke. 2001. The causes of land-
use and land-cover change: moving beyond the
myths. Global Environmental Change 11:261–269.
Laurance, W. F., A. K. Albernaz, G. Schroth, P. M.
Fearnside, S. Bergen, E. M. Venticinque, and C. Da
Costa. 2002. Predictors of deforestation in the Brazil-
ian Amazon. Journal of Biogeography 29:737–748.
Lee, D. 2013. CARBayes: an R Package for Bayesian
spatial modeling with conditional autoregressive
priors. Journal of Statistical Software 55:1–24.
Legendre, P. 1993. Spatial autocorrelation – Trouble or
new paradigm. Ecology 74:1659–1673.
Legendre, P., and M. J. Fortin. 1989. Spatial pattern
and ecological analysis. Vegetatio 80:107–138.
Lennon, J. J. 2000. Red-shifts and red herrings in
geographical ecology. Ecography 23:101–113.
Leroux, B. G., X. Lei, and N. Breslow. 2000. Estimation
of disease rates in small areas: a new mixed model
for spatial dependence. Pages 179–191 in M. E.
Halloran and D. Berry, editors. Statistical models in
epidemiology, the environment, and clinical trials.
Springer, New York, New York, USA.

pez-Carr, D. 2008. Population and deforestation:
Lo
why rural migration matters. Progress in Human
Geography 33:355–378.

pez-Carr, D., and J. Burgdorfer. 2013. Deforestation
Lo
drivers: population, migration, and tropical land
use. Environment: Science and Policy for Sustain-
able Development 55:3–11.
Mainwaring, S. 2006. The crisis of representation in the
Andes. Journal of Democracy 17:13–27.
Montenegro, E. C., D. M. V. Galvis, G. G. Garc
ıa, M. C.
Garc
ıa D
avila, M. F. Ordo
n
~ ez Castro, L. K. Vergara,
A. M. Pascagaza, J. C. Rubiano, and P. G. Rodriguez.
2011. Memoria T
ecnica de la Cuantificacio
n de la
Deforestacio
n Histo
rica Nacional Escalas Gruesa y
Fina. Instituto de Hidrolog
ıa, Meteorolog
ıa, y Estu-
dios Ambientales, Bogot
a, Colombia, USA.
NASA. 2015. Fire Information for Resource Manage-
ment System (FIRMS). http://earthdata.nasa.gov/da
ta/near-real-time-data/firms/active-fire-data
Nepstad, D., S. Schwartzman, B. Bamberger, M. San-
tilli, D. Ray, P. Schlesinger, P. Lefebvre, A. Alencar,
E. Prinz, and G. Fiske. 2006. Inhibition of Amazon
deforestation and fire by parks and indigenous
lands. Conservation Biology 20:65–73.
Nepstad, D., B. S. Soares-Filho, F. Merry, A. Lima,
P. Moutinho, J. Carter, M. Bowman, A. Cattaneo,
H. Rodrigues, and S. Schwartzman. 2009. The end
of deforestation in the Brazilian Amazon. Science
326:1350–1351.
Newman, M. E., K. P. McLaren, and B. S. Wilson. 2014.
Assessing deforestation and fragmentation in a
16 May 2017 ❖ Volume 8(5) ❖ Article e01824

tropical moist forest over 68 years; the impact of
roads and legal protection in the Cockpit Country,
Jamaica. Forest Ecology and Management 315:138–
152.
Nolte, C., A. Agrawal, K. M. Silvius, and B. S. Soares-
Filho. 2013. Governance regime and location influ-
ence avoided deforestation success of protected
areas in the Brazilian Amazon. Proceedings of the
National Academy of Sciences of the United States
of America 110:4956–4961.
Ord, J. K., and A. Getis. 1995. Local spatial autocorrela-
tion statistics: distributional issues and an applica-
tion. Geographical Analysis 27:286–306.
Overmars, K., G. De Koning, and A. Veldkamp. 2003.
Spatial autocorrelation in multi-scale land use
models. Ecological Modelling 164:257–270.
Perz, S. G., Y. Qiu, Y. Xia, J. Southworth, J. Sun,
M. Marsik, K. Rocha, V. Passos, D. Rojas, and
G. Alarco
n. 2013. Trans-boundary infrastructure
and land cover change: highway paving and com-
munity-level deforestation in a tri-national frontier
in the Amazon. Land Use Policy 34:27–41.
Pfaff, A. S. 1999. What drives deforestation in the
Brazilian Amazon?: evidence from satellite and
socioeconomic data. Journal of Environmental
Economics and Management 37:26–43.
Pfaff, A., J. Robalino, E. Lima, C. Sandoval, and L. D.
Herrera. 2014. Governance, location and avoided
deforestation from protected areas: Greater restric-
tions can have lower impact, due to differences in
location. World Development 55:7–20.
Pfaff, A., J. Robalino, R. Walker, S. Aldrich, M. Caldas,
E. Reis, S. Perz, C. Bohrer, E. Arima, and W. Lau-
rance. 2007. Road investments, spatial spillovers,
and deforestation in the Brazilian Amazon. Journal
of Regional Science 47:109–123.
Pinheiro, J., D. Bates, S. DebRoy, and D. Sarkar. 2011. R
Development Core Team. 2010. nlme: linear and
nonlinear mixed effects models. R package version
3.1-97. R Foundation for Statistical Computing,
Vienna, Austria.
Plummer, M. 2003. JAGS: A program for analysis of
Bayesian graphical models using Gibbs sampling.
Proceedings of the 3rd International Workshop on
Distributed Statistical Computing. Technische
Universit€at Wien, Vienna, Austria.
Pokorny, B., W. de Jong, J. Godar, P. Pacheco, and
J. Johnson. 2013. From large to small: reorienting
rural development policies in response to climate
change, food security and poverty. Forest Policy
and Economics 36:52–59.
Portocarrero-Aya, M., G. Corzo, A. Diaz-Pulido, M. F.
Gonz
alez, M. Longo, L. Mesa, A. Paz, W. Ram
ırez,
and O. L. Hern
andez-Manrique. 2014. Systematic
conservation assessment for most of the Colombian
❖ www.esajournals.org
METS ET AL.
territory as a strategy for effective biodiversity con-
servation. Natural Resources 5:981.
R Development Core Team. 2008. R: a language and
environment for statistical computing. R Founda-
tion for Statistical Computing, Vienna, Austria.
Rodrigues, A. S., R. M. Ewers, L. Parry, C. Souza,
A. Ver
ıssimo, and A. Balmford. 2009. Boom-
and-bust development patterns across the Amazon
deforestation frontier. Science 324:1435–1437.
Rodr
ıguez, N., D. Armenteras, and J. R. Alumbreros.
2013a. Land use and land cover change in the
Colombian Andes: dynamics and future scenarios.
Journal of Land Use Science 8:154–174.
Rodr
ıguez, N., D. Armenteras, R. Molowny-Horas,
and J. Retana. 2012. Patterns and trends of forest loss
in the Colombian Guyana. Biotropica 44:123–132.
Rodr
ıguez, N., D. Armenteras, and J. Retana. 2013b.
Effectiveness of protected areas in the Colombian
Andes: deforestation, fire and land-use changes.
Regional Environmental Change 13:423–435.
Rudel, T. K. 2007. Changing agents of deforestation:
from state-initiated to enterprise driven processes,
1970–2000. Land Use Policy 24:35–41.
Rudel, T. K., R. Defries, G. P. Asner, and W. F.
Laurance. 2009. Changing drivers of deforestation
and new opportunities for conservation. Conserva-
tion Biology 23:1396–1405.
Rudel, T., and J. Roper. 1997. The paths to rain forest
destruction: crossnational patterns of tropical
deforestation, 1975–1990. World Development 25:
53–65.
Sanchez-Cuervo, A. M., and T. M. Aide. 2013. Identify-
ing hotspots of deforestation and reforestation in
Colombia (2001–2010): implications for protected
areas. Ecosphere 4:art143.
S

anchez-Cuervo, A. M., and T. M. Aide. 2013. Conse-
quences of the armed conflict, forced human dis-
placement, and land abandonment on forest cover
change in Colombia: a multi-scaled analysis.
Ecosystems 16:1052–1070.
S

anchez-Cuervo, A. M., T. M. Aide, M. L. Clark, and
A. Etter. 2012. Land cover change in Colombia:
surprising forest recovery trends between 2001 and
2010. PLoS ONE 7:e43943.
Scherr, S. J. 2000. A downward spiral? Research evi-
dence on the relationship between poverty and nat-
ural resource degradation. Food Policy 25:479–498.
Schuurman, F. J. 1978. From resource frontier to
periphery agricultural colonization east of the
Andes. Tijdschrift voor economische en sociale
geografie 69:95–104.
Seto, K. C., R. S
anchez-Rodr
ıguez, and M. Fragkias.
2010. The new geography of contemporary urban-
ization and the environment. Annual Review of
Environment and Resources 35:167–194.
17 May 2017 ❖ Volume 8(5) ❖ Article e01824
 METS ET AL.

Soares-Filho, B., A. Alencar, D. Nepstad, G. Cerqueira,
V. Diaz, M. del Carmen, S. Rivero, L. Solo
rzano,
and E. Voll. 2004. Simulating the response of land-
cover changes to road paving and governance along
a major Amazon highway: the Santar
em-Cuiab
a UNFCCC [United Nations Framework Convention on
corridor. Global Change Biology 10:745–764.
Soares-Filho, B., P. Moutinho, D. Nepstad, A. Ander-
son, H. Rodrigues, R. Garcia, L. Dietzsch, F. Merry,
M. Bowman, and L. Hissa. 2010. Role of Brazilian
Amazon protected areas in climate change mitiga-
tion. Proceedings of the National Academy of
Sciences of the United States of America 107:
10821–10826.
Southgate, D. 1990. The causes of land degradation
along” spontaneously” expanding agricultural fron-
tiers in the Third World. Land Economics 66:93–101.
Su, Y. S., and M. Yajima. 2015. R2jags: a package for
running jags from R. http://cran.r-project.org/packa
ge=R2jags
Telford, R., and H. Birks. 2005. The secret assumption
of transfer functions: problems with spatial auto-
correlation in evaluating model performance. Qua-
ternary Science Reviews 24:2173–2179.
Climate Change]. 2015. Adoption of the Paris
Agreement -/CP21. UNFCCC, Paris, France.
UNODC, Government of Colombia. 2006. Colombia:
Coca cultivation survey. United Nations Office
on Drugs and Crime, Bogot
a, Colombia. http://
www.unodc.org/pdf/andean/Colombia_coca_survey_
2005_eng.pdf
Yu, L., N. Wang, and X. Meng. 2005. Real-time forest
fire detection with wireless sensor networks. Pages
1214–1217 in H. Zhou, editor. Proceedings of Inter-
national Conference on Wireless Communications,
Networking and Mobile Computing, Wuhan, China,
September 23–26, 2005. Institute of Electrical & Elec-
tronics Engineers, New York, New York, USA.

SUPPORTING INFORMATION
Additional Supporting Information may be found online at: http://onlinelibrary.wiley.com/doi/10.1002/ecs2.
1824/full

❖ www.esajournals.org 18 May 2017 ❖ Volume 8(5) ❖ Article e01824